Domain adaptation for wireless sensing

Abstract

Certain aspects of the present disclosure provide techniques for domain adaptation. An input tensor comprising channel state information (CSI) for a wireless signal is determined, where each channel in the input tensor corresponds to a respective degree of freedom (DoF) in the wireless signal. A domain-adapted tensor is generated by processing the input tensor using a domain-adaptation network comprising, for each respective DoF in the wireless signal, a respective convolution path. The domain-adapted tensor is provided to a neural network trained for position estimation.

What is claimed is: 
     
         1 . A method, comprising:
 determining an input tensor comprising channel state information (CSI) for a wireless signal, wherein each channel in the input tensor corresponds to a respective degree of freedom (DoF) in the wireless signal;   generating a domain-adapted tensor by processing the input tensor using a domain-adaptation network comprising, for each respective DoF in the wireless signal, a respective convolution path; and   providing the domain-adapted tensor to a neural network trained for position estimation.   
     
     
         2 . The method of  claim 1 , wherein each DoF in the wireless signal corresponds to a transmitting antenna and receiving antenna pair for the wireless signal. 
     
     
         3 . The method of  claim 2 , wherein processing the input tensor using the domain-adaptation network comprises:
 generating a channel tensor by applying a global pooling operation to the input tensor;   generating a plurality of vectors by processing each respective channel in the channel tensor using a respective convolution path;   generating a DoF matrix by concatenating the plurality of vectors;   generating a channel weighting vector by processing the channel tensor using a softmax activation function; and   generating an attention matrix by multiplying the DoF matrix and the channel weighting vector.   
     
     
         4 . The method of  claim 3 , further comprising generating a mask matrix by processing the attention matrix using a non-linear transformation function. 
     
     
         5 . The method of  claim 4 , further comprising generating a residual tensor by performing element-wise multiplication between the mask matrix and the input tensor. 
     
     
         6 . The method of  claim 5 , further comprising:
 processing the input tensor using a linear transformation to generate a scaled input tensor; and   generating the domain-adapted tensor by adding the residual tensor and the scaled input tensor.   
     
     
         7 . The method of  claim 2 , wherein determining the input tensor comprises identifying CSI associated with each DoF based on a reference signal pattern used to transmit the wireless signal. 
     
     
         8 . The method of  claim 1 , wherein providing the domain-adapted tensor to the neural network comprises providing the domain-adapted tensor to one or more feature extraction layers of the neural network. 
     
     
         9 . The method of  claim 1 , wherein:
 a first spatial dimension in the input tensor corresponds to a number of subcarriers per orthogonal frequency-division multiplexing (OFDM) symbol in the wireless signal; and   a second spatial dimension in the input tensor corresponds to a number of OFDM symbols over time in the wireless signal.   
     
     
         10 . The method of  claim 1 , further comprising:
 applying one or more random affine transformations to a training data batch to produce a transformed training data batch; and   training each respective convolution path based on the training data batch and the transformed training data batch.   
     
     
         11 . The method of  claim 10 , wherein applying the one or more random affine transformations is defined as Y′=UY+V, where:
 Y′ is the transformed training data batch; 
 Y is the training data batch; and 
 U and V are sampled from a distribution. 
 
     
     
         12 . A processing system, comprising:
 a memory comprising computer-executable instructions; and   one or more processors configured to execute the computer-executable instructions and cause the processing system to perform an operation comprising:
 determining an input tensor comprising channel state information (CSI) for a wireless signal, wherein each channel in the input tensor corresponds to a respective degree of freedom (DoF) in the wireless signal; 
 generating a domain-adapted tensor by processing the input tensor using a domain-adaptation network comprising, for each respective DoF in the wireless signal, a respective convolution path; and 
 providing the domain-adapted tensor to a neural network trained for position estimation. 
   
     
     
         13 . The processing system of  claim 12 , wherein each DoF in the wireless signal corresponds to a transmitting antenna and receiving antenna pair for the wireless signal. 
     
     
         14 . The processing system of  claim 13 , wherein processing the input tensor using the domain-adaptation network comprises:
 generating a channel tensor by applying a global pooling operation to the input tensor;   generating a plurality of vectors by processing each respective channel in the channel tensor using a respective convolution path;   generating a DoF matrix by concatenating the plurality of vectors;   generating a channel weighting vector by processing the channel tensor using a softmax activation function; and   generating an attention matrix by multiplying the DoF matrix and the channel weighting vector.   
     
     
         15 . The processing system of  claim 14 , the operation further comprising:
 generating a mask matrix by processing the attention matrix using a non-linear transformation function;   generating a residual tensor by performing element-wise multiplication between the mask matrix and the input tensor;   processing the input tensor using a linear transformation to generate a scaled input tensor; and   generating the domain-adapted tensor by adding the residual tensor and the scaled input tensor.   
     
     
         16 . The processing system of  claim 13 , wherein determining the input tensor comprises identifying CSI associated with each DoF based on a reference signal pattern used to transmit the wireless signal. 
     
     
         17 . The processing system of  claim 12 , wherein providing the domain-adapted tensor to the neural network comprises providing the domain-adapted tensor to one or more feature extraction layers of the neural network. 
     
     
         18 . The processing system of  claim 12 , wherein:
 a first spatial dimension in the input tensor corresponds to a number of subcarriers per orthogonal frequency-division multiplexing (OFDM) symbol in the wireless signal; and   a second spatial dimension in the input tensor corresponds to a number of OFDM symbols over time in the wireless signal.   
     
     
         19 . The processing system of  claim 12 , the operation further comprising:
 applying one or more random affine transformations to a training data batch to produce a transformed training data batch; and   training each respective convolution path based on the training data batch and the transformed training data batch.   
     
     
         20 . The processing system of  claim 19 , wherein applying the one or more random affine transformations is defined as Y′=UY+V, where:
 Y′ is the transformed training data batch; 
 Y is the training data batch; and 
 U and V are sampled from a distribution. 
 
     
     
         21 . A non-transitory computer-readable medium comprising computer-executable instructions that, when executed by one or more processors of a processing system, cause the processing system to perform an operation comprising:
 determining an input tensor comprising channel state information (CSI) for a wireless signal, wherein each channel in the input tensor corresponds to a respective degree of freedom (DoF) in the wireless signal;   generating a domain-adapted tensor by processing the input tensor using a domain-adaptation network comprising, for each respective DoF in the wireless signal, a respective convolution path; and   providing the domain-adapted tensor to a neural network trained for position estimation.   
     
     
         22 . The non-transitory computer-readable medium of  claim 21 , wherein each DoF in the wireless signal corresponds to a transmitting antenna and receiving antenna pair for the wireless signal. 
     
     
         23 . The non-transitory computer-readable medium of  claim 22 , wherein processing the input tensor using the domain-adaptation network comprises:
 generating a channel tensor by applying a global pooling operation to the input tensor;   generating a plurality of vectors by processing each respective channel in the channel tensor using a respective convolution path;   generating a DoF matrix by concatenating the plurality of vectors;   generating a channel weighting vector by processing the channel tensor using a softmax activation function; and   generating an attention matrix by multiplying the DoF matrix and the channel weighting vector.   
     
     
         24 . The non-transitory computer-readable medium of  claim 23 , the operation further comprising:
 generating a mask matrix by processing the attention matrix using a non-linear transformation function;   generating a residual tensor by performing element-wise multiplication between the mask matrix and the input tensor;   processing the input tensor using a linear transformation to generate a scaled input tensor; and   generating the domain-adapted tensor by adding the residual tensor and the scaled input tensor.   
     
     
         25 . The non-transitory computer-readable medium of  claim 22 , wherein determining the input tensor comprises identifying CSI associated with each DoF based on a reference signal pattern used to transmit the wireless signal. 
     
     
         26 . The non-transitory computer-readable medium of  claim 21 , wherein providing the domain-adapted tensor to the neural network comprises providing the domain-adapted tensor to one or more feature extraction layers of the neural network. 
     
     
         27 . The non-transitory computer-readable medium of  claim 21 , wherein:
 a first spatial dimension in the input tensor corresponds to a number of subcarriers per orthogonal frequency-division multiplexing (OFDM) symbol in the wireless signal; and   a second spatial dimension in the input tensor corresponds to a number of OFDM symbols over time in the wireless signal.   
     
     
         28 . The non-transitory computer-readable medium of  claim 21 , further comprising:
 applying one or more random affine transformations to a training data batch to produce a transformed training data batch; and   training each respective convolution path based on the training data batch and the transformed training data batch.   
     
     
         29 . The non-transitory computer-readable medium of  claim 28 , wherein applying the one or more random affine transformations is defined as Y′=UY+V, where:
 Y′ is the transformed training data batch; 
 Y is the training data batch; and 
 U and V are sampled from a distribution. 
 
     
     
         30 . A processing system, comprising:
 means for determining an input tensor comprising channel state information (CSI) for a wireless signal, wherein each channel in the input tensor corresponds to a respective degree of freedom (DoF) in the wireless signal;   means for generating a domain-adapted tensor by processing the input tensor using a domain-adaptation network comprising, for each respective DoF in the wireless signal, a respective convolution path; and   means for providing the domain-adapted tensor to a neural network trained for position estimation.